AbstractStunting is recognised as a major public health problem in

Rwanda. We therefore aimed to study the demographic, socio-eco-nomic and environmental factors determining the spatial patternof stunting. A cross-sectional study using the data from the 2014-2015 Rwanda Demographic and Health Survey and environmental

data from external geospatial datasets were conducted. The studypopulation was children less than two years old with their moth-ers. A multivariate linear regression model was used to estimatethe effects of demographic, socio-economic and biophysical fac-tors and a proxy measure of aflatoxins exposure on height-for-age.Also, a spatial prediction map of height-for-age to examine thestunting pattern was produced. It was found that age of child,height of mother, secondary education and higher, a child beingmale and birth weight were associated with height-for-age. Afteradjusting for demographic and socioeconomic factors, elevationand being served by a rural market were also significantly associ-ated with low height-for-age in children. The spatial predictionmap revealed the variability of height-for-age at the cluster-levelthat was lost when the levels are aggregated at the district level.No associations with height-for-age were found for exclusivebreastfeeding, use of deworming tablets, improved water sourceand improved sanitation in the study population. In addition to thechild and mother factors known to determine height-for-age, ourstudy confirms the influence of environmental factors in determin-ing the height-of-age of children in Rwanda. A consideration ofthe environmental drivers of anthropometric status is crucial tohave a holistic approach to reduce stunting.

IntroductionStunting among children remains a public health concern in

Sub-Saharan Africa. Eastern and Central Africa have the higheststunting levels at 36.7% and 32.5%, respectively (FAO et al.,2017). Stunting results in decreased motor and cognitive develop-ment, impaired immunity and low education attainment (Deweyand Begum, 2011). In adulthood, stunting leads to lower economicproductivity, increased risk of chronic diseases and lower off-spring birth weight (Victora et al., 2008). Stunting increasesmarkedly in the first two years of life especially during the com-plementary feeding period (Black et al., 2008; Victora et al.,2010).

Stunting is the outcome of a complexity of factors. A pro-longed lack of adequate and nutritious diet, coupled with recurringinfections are direct causes of stunting. Child-related factors suchas age, lower birth weight, lack of exclusive breastfeeding, andinadequate complementary feeding practices also increase the riskof stunting. Stunting is also influenced by maternal factors (e.g.short maternal stature, lower education, short birth spacing, poorhealth before conception), and by household and living character-istics (e.g. low income, inadequate water and sanitation and loweraccess to health services) (Stewart et al., 2013). The biophysical

Correspondence: Vestine Uwiringiyimana, Faculty of Geo-InformationScience and Earth Observation (ITC), University of Twente,Hengelosestraat 99 (359), 7514 AE Enschede, The Netherlands.E-mail: v.uwiringiyimana@utwente.nl

Key words: Stunting; Environment; Height-for-age; Spatial pattern;Aflatoxin; Rwanda.

Acknowledgements: the authors thank the Demographic and HealthSurvey (DHS) Programme for granting them permission to use theRwanda 2015 DHS data. The authors would like to also thank Mr. MarcManyifika and Dr. Elias Nyandwi for their assistance in getting the spa-tial data used in the study.

Contributions: VU formulated the research question, designed thestudy, conducted the statistical analysis, interpreted the findings andwrote the manuscript. SA contributed to the interpretation of findings,the writing and critical revision of the manuscript; AV formulated theresearch question, designed the study, interpreted the findings, criticallyrevised the manuscript. All authors read and approved the final versionof the manuscript.

Conflict of interest: the authors declare no potential conflict of interest.

Funding: this work was supported by the Nuffic fundedNICHE/RWA/071 project.

Received for publication: 3 October 2019.Revision received: 1 November 2019.Accepted for publication: 1 November 2019.

©Copyright: the Author(s), 2019Licensee PAGEPress, ItalyGeospatial Health 2019; 14:820doi:10.4081/gh.2019.820

This article is distributed under the terms of the Creative CommonsAttribution Noncommercial License (CC BY-NC 4.0) which permits anynoncommercial use, distribution, and reproduction in any medium, pro-vided the original author(s) and source are credited.

Stunting spatial pattern in Rwanda: An examination of the demographic,socio-economic and environmental determinantsVestine Uwiringiyimana,1,2 Antonie Veldkamp,1 Sherif Amer1

1Faculty of Geo-Information Science and Earth Observation (ITC), University of Twente, Enschede, The Netherlands; 2Department of Food Science and Technology, College of Agriculture Animal Science and Veterinary Medicine, University of Rwanda, Kigali, Rwanda

[Geospatial Health 2019; 14:820] [page 329]

Geospatial Health 2019; volume 14:820

Non-co

mmercial

use o

nly

[page 330] [Geospatial Health 2019; 14:820]

environment influences household food security and thereforedirectly or indirectly affects dietary diversity. Biophysical factorsoften studied include elevation, rainfall, temperature, relativehumidity, topography, soil type, aridity, distance to urban areas andmarket access (de Sherbinin, 2011; Nikoi and Anthamatten, 2013).

Apart from nutrition, demographic, socio-economic and bio-physical factors, mycotoxins exposure has also been linked tostunting. In many developing countries exposure to mycotoxinsoccurs via the food supply chain (Udomkun et al., 2017).Mycotoxins are toxic secondary metabolites produced by a groupof fungal species that grow on cereals, nuts and legumes before orafter harvest, especially under warm and humid conditions (Pitt,2000; Cotty and Jaime-Garcia, 2007). Because many developingcountries rely on subsistence farming and informal local markets,mycotoxins control is sometimes non-existent (Gbashi et al.,2018). Child exposure to mycotoxins, especially during the wean-ing period, is harmful as it impairs immunity and linear growth(Shephard, 2008; Khlangwiset et al., 2011; Vilcins et al., 2018).Among mycotoxins, exposure to aflatoxins and fumonisins hasbeen consistently linked to stunting in children (Wild and Gong,2010; Amuzie et al., 2016). Maize and peanut flours, often givento children as part of complementary feeding, are prime pathwaysfor aflatoxins and fumonisins exposure (Udomkun et al., 2017).

In the past decade, Rwanda has made remarkable progress inimproving food security (MINAGRI et al., 2016), access to healthcare (Lu et al., 2016), water and sanitation. This has led to reducedchildhood undernutrition, especially wasting (NISR et al., 2015).In 2013, the country set a target of reducing child stunting to24.5% by 2018 (MINALOC et al., 2014). Stunting howeverremains high with 38% of children under five being classified asstunted (NISR et al., 2015). It is estimated that Rwanda loses11.5% of its gross domestic product due to child undernutrition(AUC and NEPAD, 2013). Recognising the complexity of stunt-ing, Rwanda adopted a multi-sectoral approach focusing on earlychildhood development, improving water and sanitation as well asincreasing social protection (MIGEPROF, 2018). On the otherhand, the underlying factors leading to high levels of stunting insome regions in Rwanda remain unclear (MIGEPROF, 2018).Also, the contribution of mycotoxins exposure to the high levels ofobserved stunting in Rwanda remains a subject that is not wellinvestigated. A lack of national level data on mycotoxins occur-rence represents a serious challenge to mycotoxin research inRwanda and in other developing countries (Gbashi et al., 2018).

To identify the local determinants of stunting, we conductedresearch in the northern region of Rwanda, which in 2010 wasamong the districts with the highest level of stunting(Uwiringiyimana et al., 2018). Follow-up research examined theimpact of aflatoxin exposure on children’s height-for-age, andestablished that exposure to aflatoxins via complementary flourswas indeed associated with low child height-for-age(Uwiringiyimana, Ocke, Amer, Veldkamp, unpublished data). Theanalysis presented here takes a holistic approach as it not only con-siders the usual demographic and socio-economic factors known toinfluence stunting but also incorporates biophysical factors andexposure to mycotoxins.

Spatially explicit nationally representative data on mycotoxinsin maize and peanut flours however do not exist in Rwanda. Toovercome this lack of data, we developed a proxy for mycotoxinexposure which is based on the food supply chain of maize andpeanut flours. Nationally, most households (70%) get their com-plementary flours from markets (MINAGRI et al., 2016). In a local

study conducted in the northern region of Rwanda(Uwiringiyimana, Ocke, Amer, Veldkamp, unpublished data),nearly all households surveyed (99%) obtained their peanut flourfrom a market. Given that aflatoxin production will generallyincrease along the food supply chain and during storage (Hell andMutegi, 2011), we hypothesised that household clusters that aresupplied by markets located at the lower end of the food supplychain would stand an increased risk of mycotoxin exposure. Theobjectives of this study are threefold. First, to examine the demo-graphic and socio-economic factors associated with the observedheight-for-age variability in Rwanda. Second, controlling for thedemographic and socio-economic factors, to identify the biophys-ical factors affecting height-for-age variability. Third, adjusting forthe previous factors, to investigate if obtaining flour from thelower end of the food supply chain was associated with lowerchildhood height-for-age.

Materials and Methods

Study areaRwanda is a country located in East-Africa with Tanzania to

the East, Uganda to the North, Congo to the West and Burundi tothe South. About 61.5% of Rwanda lies above 1,500 meters abovethe mean sea level, giving the country a temperate climate(Mendelsohn et al., 2016). Rwanda has four provinces, 30 districts,416 sectors, 2,148 cells and 14,837 villages. The population isabout 11 million, among which 1,5 million are under five (NISRand MINECOFIN, 2012). Figure 1 shows the spatial variability ofstunting among the 30 districts of Rwanda. This study was basedat the cluster level (containing between 100 and 300 households).

Study sampleThe survey was conducted in a two-stage sample design based

on demographic and socio-economic data from the 2014-15 col-lected from the Rwanda Demographic and Health Survey (DHS)(NISR et al., 2015). At the first stage, a total of 492 householdclusters were selected within enumeration areas representing all 30districts of Rwanda, 113 located in urban areas, and 379 in ruralareas. For survey participation, 20 to 30 households were random-ly selected from each cluster. The second stage involved systemat-ic sampling of 26 households within each cluster, making up asample of 12,792 households at the national level. From a sub-sample consisting of 50% of the households, indices of anthropo-metric status (child height and weight), individual mother andhousehold characteristics were recorded for 3,813 children lessthan five years old. At the end of the survey, 1,467 children agedless than two years had valid height-for-age measurements,matched with their mothers’ characteristics. All the demographicand socio-economic variables used were extracted from the 2014-15 DHS database for Rwanda, taking into account the surveyweights, the variables definition and the denominators (Rutsteinand Rojas, 2006).

Demographic and socio-economic dataThe outcome variable was the height-for-age z-scores (HAZ)

of children less than two years. Height-for-age is a measure ofstunting or low height-for-age which indicates a prolonged nutrientdeficit in children resulting in linear growth retardation (de Onis

Article

Non-co

mmercial

use o

nly

and Branca, 2016). A z-score less than -2 standard deviations(SDs) below the median of the World Health Organization (WHO)reference population (Onis, 2007) defines stunting. The indepen-dent demographic variables used were: i) age, sex, birth weight,exclusive breastfeeding, use of deworming tablets in the last sixmonths, and the presence of diarrhoea in the two weeks precedingthe survey for a child; ii) height, body mass index (BMI), highesteducation level, preceding birth interval (the difference in monthsbetween the current birth and the previous birth, counting twins asone birth), education level (no education, primary education, sec-ondary education or higher) for a mother.

The socio-economic variables used were: i) type of residence,improved source of water, improved sanitation or toilet facility; ii)access to health care facility; iii) wealth index (a composite mea-sure of a household’s economic status, estimated from ownershipof selected assets such as television, bicycles, type of residenceincluding water access/sanitation facilities and access to healthcare facility; NISR et al., 2015).

Environmental dataThe biophysical variables considered in the study were eleva-

tion, slope, temperature, rainfall and relative humidity. Table 1 dis-plays the description of the environmental data. The elevation wasproduced from a digital elevation map (DEM) created by theRwanda National Land Use and Development Master Plan Projectthrough the Rwanda Natural Resources Authority. The slope wascalculated from the DEM. The climatic factors temperature, rain-fall, and relative humidity were measured from a network of 183meteorological stations throughout Rwanda (Nyandwi et al.,2016). The annual average of these factors was measured for moststations using 60 years records starting from 1950 to 2010 by theRwanda Meteorological Agency. After, the averages were interpo-lated for the whole country using thin-plate smoothing spline algo-rithm and saved as raster datasets with a 10-m2 cell size (Nyandwiet al., 2016).

Article

Table 1. Description of environmental variables.

Data type Variable Description

Topographic Elevation and slope Created from a DEM and further used to generate the slope using ArcGIS. Both raster files were created at a cell size of 10 m2

Climatic Temperature, rainfall, and relative humidity Annual means from meteorological weather stations, interpolated and resampled to a cell size of 10 m2

DEM, digital elevation map.

Figure 1. Stunting prevalence per district in Rwanda in 2015. Based on Demographic and Health Survey, 2015 (Rwanda location inAfrica shown besides).

[Geospatial Health 2019; 14:820] [page 331]

Non-co

mmercial

use o

nly

Merging environmental data and Demographic andHealth Survey data

To merge the environmental data with DHS data, each bio-physical factor was extracted per household cluster using zonalstatistics in ArcGIS Desktop, release 10.4 (ESRI, Redlands, CA,USA). Consideration was made, however, by taking into accountthe geographical displacement of the cluster done by DHS forevery household survey (Burgert et al., 2013). For protecting theconfidentiality of survey participants, DHS displaces the globalpositioning system (GPS) coordinates of urban clusters up to twokm, and of rural clusters up to 5 km. Thus, before extracting thebiophysical factors per cluster, we created a buffer of 5 km aroundall rural clusters and a buffer of 2 km around all urban clusters. Themean value of each environmental variable was extracted using thezonal statistics tool in ArcGIS.

Classification of household clustersAs the environmental conditions in the tropics are known to be

very conducive to the production of mycotoxins (Gbashi et al.,2018), and given that aflatoxin contamination increases along thefood supply chain due to poor post-harvest management (Kamikaet al., 2016), we considered household clusters that would obtaintheir flour from markets at the lower end of the food supply chainas being at risk of exposure to higher aflatoxin levels from comple-mentary flour compared to household clusters that obtain theirflour at the start of the food supply chain. Because the food supplychain is a hierarchy from main markets to regional and local mar-kets, we classified the clusters as follows: household clustersserved by an urban market were considered as obtaining their flourfrom markets at the top of the food supply chain (least risk of afla-toxin exposure), while clusters exclusively served by a rural mar-ket were considered as obtaining their flour from markets at theintermediate level (increased risk of exposure) and clusters obtain-ing flour from local shops at the end of the food supply chain(highest risk of exposure). Based on this situation, the road net-work of Rwanda, the market locations and the boundaries of urbanareas were used to classify the clusters into data layers.

The national road network, produced by the Rwanda Land

Management and Use Authority (RLMUA) within the land cadas-tral project in 2009-2010 categorises roads into national, districtand other roads. The existing boundary of urban areas in Rwandawas also delineated by RLMUA during its land demarcation pro-ject for the production of urban master plans of Rwandan districts.The market location was produced by the National Institute ofStatistics of Rwanda (NISR) by collecting information on the loca-tion of each market on the ground. A summary of the process fol-lowed to compute the distance variables is shown in Figure 2. Wecomputed the distance from each household cluster to the closesturban or rural market in ArcGIS.

To prepare the data for network analysis, a road layer was firstbuilt followed by a market layer with the existing boundary ofurban areas. The markets were extracted and categorised as urbanor rural taking into account that main markets in urban areas sup-plying flour to rural markets. Then, two distance variables werecalculated from each household cluster to the closest urban andrural markets using the network analyst tool.

Because of the geographic displacement of the household clus-ters done by DHS to protect the confidentiality of survey respon-dents, using a continuous distance variable from a household clus-ter to any closest facility could introduce a bias in the measure-ments computed. To reduce this bias, categorising the distanceaccording Burgert et al. (2013) we used two categories for distanceto market (5 and 10 km). These two cut-offs were chosen to con-sider the walkable distance from a household cluster to the closestmarket, and also because they take into account the geographicaldisplacement of both the urban and the rural household clusters.The distance variables were used to classify the household clustersinto the three categories depending on where they would obtaintheir flour, all within a 5 or 10 km radius: i) household clustersserved by an urban market; ii) household clusters exclusivelyserved by a rural market; and iii) household clusters neither servedby an urban market nor a rural one.

Statistical analysisAll variables were imported into SPSS Statistics for Windows,

version 24.0 (IBM Corp., Armonk, NY, USA) for statistical analy-

Article

Figure 2. Flowchart explaining the calculation of distance to markets.

[page 332] [Geospatial Health 2019; 14:820]

Non-co

mmercial

use o

nly

sis. Descriptive statistics were run on all variables. Mean, medianand SD were reported for continuous variables. Percentages weredisplayed for discrete ones. Maps of the independent variablesaggregated per household cluster were produced to visualise theirspatial pattern. For the spatial pattern of categorical variables, thepercentage levels were displayed. For continuous variables, suchas mothers BMIs, classes of BMI indicating underweight, normalweight, overweight and obesity, were used to visualise the spatialpattern. Correlation between the anthropometric status and theindependent variables were ran on all variables to study the linearrelationship with height-for-age. A hierarchical stepwise linearregression model was applied to determine the factors that weresignificant HAZ determinants. Three hierarchical stepwise modelswere run, with demographic and socio-economic factors that affectheight-for-age entered first. The second block consisted of envi-ronmental variables and in the third block, the household clustersclassification variables, based on the type of market that servedeach cluster. The independent variables retained in the final modelwere all significant at the P≤0.05 level. Given that most regressionanalyses have been criticised of overlooking or not correcting forthe spatial dependence which biases the results of the regressionmodels (Voss et al., 2006; de Sherbinin, 2011), we checked forpossible spatial autocorrelation in our analyses. Thus, the results of

the final model were validated by visualising the pattern of theresiduals in ArcGIS, and running a spatial autocorrelation analysisusing Global Moran’s I to assess the non-spatial clustering of themodel residuals. Spatial autocorrelation indicates that the errors inthe model are not independent (Voss et al., 2006). When spatialautocorrelation is present, it reduces the standard errors of the esti-mates, increases the t-values and reduces the P-value thereby lead-ing to a bias in the model (Voss et al., 2006). Moran’s I is similarto Pearson’s correlation coefficient, and has values ranging from -1 to 1, with positive values indicating high values surrounded byhigh values and, negative results showing spatial randomness. Aprediction map based on the model predicted height-for-age z-scores per household cluster and per district was produced usingnatural breaks.

ResultsThe descriptive characteristics of the study population is pre-

sented in Table 2, while Figure 3 displays the spatial variabilityexisting within the demographic and socio-economic variables percluster. The high percentage of stunting is more common in theNorth, West and south-eastern part of Rwanda and varies from 0%

Article

[Geospatial Health 2019; 14:820] [page 333]

Figure 3. Prevalence rate of Demographic and Health Survey variables per cluster. A) Prevalence rate of stunting; B) percentage of exclu-sive breastfeeding; C) secondary and other higher education of mother; D) mothers’ body mass index (BMI) classes per cluster.

Non-co

mmercial

use o

nly

to 100% in some clusters (Figure 3A). Exclusive breastfeeding ofchildren less than six months was generally below 75% (Figure3B). As expected, mothers with secondary education and highermostly lived in Kigali (Figure 3C). On the other hand, the BMI ofmothers were uniformly spread across the country (Figure 3D).

As shown in Figure 4, the percentage of children receivingdeworming tablets in the last six months (Figure 4A) was higheracross the country as opposed to the prevalence of diarrhoea(Figure 4B). The use of non-improved water source (Figure 4C) inthe households was more pronounced in the East, North and someregions in the South. On the other hand, the use of non-improvedsanitation (Figure 4D) was evenly spread across the four provincesof Rwanda.

Table 3 shows the results from the household cluster classifi-cation. Household clusters that were served by an urban marketwithin a 5 and a 10-km radius were 21% and 32%, respectively.Household clusters served only by a rural market were 39% and61% within the 5 and 10 km, respectively. Household clustersserved by neither an urban nor a rural market within 5 km were40% as opposed to 7% within 10 km.

Figure 5 shows the spatial distribution of the household clus-ters as served by an urban market or a rural market within a 5 km(Figure 5A) and 10 km radius (Figure 5B). The majority of house-hold clusters that were served by a market at the top of the foodsupply chain were located in urban areas. According to ourassumption of increased aflatoxin contamination along the foodsupply chain, these household clusters were considered as clusterswith the least risk of exposure to higher aflatoxin level. Householdclusters served by markets at the intermediate level in the food sup-ply chain were mostly located in rural areas; and thus they will bemore exposed to aflatoxins compared to clusters located in theurban areas, assuming the rural markets get their supply from theurban markets. Finally, the household clusters neither served by anurban nor a rural market within a 10-km radius, which is the thirdcategory, will be potentially the most exposed to higher levels ofaflatoxins.

As can be seen in Table 4 showing the results of the hierarchi-cal linear regression model, the child’s age, mother’s height, sec-ondary education, being male, birth weight was significantly asso-ciated with height-for-age. From the biophysical factors, elevationwas a significant predictor of height-for-age in children. Beingserved by a rural market within 10 km radius was negatively asso-ciated with the height-for-age in children. The HAZ significantlydecreased as a cluster changed from being served by an urban mar-ket to being served by a rural market within a 10-km radius. Theadjusted R2 of the final model was 0.27, implying that 27% of thetotal variability in stunting can be explained by this significantempirical model.

Figure 6 shows the results of the spatial autocorrelation analy-sis with residuals on the x-axis and the lagged residuals on the y-axis. The lagged residuals represent the sum of weighted residualsof neighbouring household clusters indicating that there is no spa-tial autocorrelation among residuals, thereby confirming theirindependence. Figure 7B shows the spatial pattern of the predictedvalues of HAZ at the cluster-level in Rwanda, based on the modelpresented in Table 4, with the same mean HAZ aggregated to thedistrict level. Overall, the central region of Rwanda has normalheight-for-age values in contrast to the northern and the westernregions that have mostly low height-for-age values. The aggrega-tion to the district level, however, overshadows the clusters withhigher HAZ, which are surrounded by clusters with lower HAZ.

As seen in Figure 7A, the same is observed from the comparisonwith the commonly reported stunting prevalence per district.Compared with Figure 1, the spatial variability in HAZ is lost dueto data aggregation at the district levels.

Article

Table 3. Household clusters classification.

Variable Sample size (%) 5 km 10 kmDistance categories

Household clusters served by an urban market 106 (21) 158 (32)Household clusters exclusively served by a rural market 90 (39) 300 (61)Household clusters served by neither an urban nor a rural market 196 (40) 34 (7)

Table 2. Descriptive statistics of dependent and independent vari-ables used in the study.

Variable Mean SD No. of clusters %Continuous

Height-for-age -1.2 1.5 1,467 Child age (months) 11 7.0 1,514 Child birth weight (kg) 3.3 0.6 1,509 Mother’s BMI (kg/m2) 23.1 3.5 1,509 Mother’s height (cm) 156.9 6.1 1,509 Birth index (months) 11.4 6.6 3,122 Wealth index -0.1 0.8 3,122 Elevation (m) 1702 348.0 Slope (degrees) 11.7 10.9 Rainfall (mm) 1208 259.0 Relative humidity 73.9 3.6 Categorical

Stunting 1,467 30.2Sex 3,122 Female 50.1 Male 49.9Exclusive breastfeeding 2,972 Yes 25.3Mother’s education 3,122 No education 13.3 Primary education 71.8 Secondary education 12.5 Higher 2.4Improved water source* 3,074 Yes 72.5Improved sanitation° 3,074 Yes 70.2Deworming tablets use in last six months 3,116 Yes 48.6Diarrhoea in last two weeks 3,122 Yes 17.0Data source: Demographic and Health Survey, Rwanda 2014-2015 and authors’ own calculations. SD, stan-dard deviation; BMI, body mass index. *includes piped water, public tap/standpipe, tube well, borehole,protected well, protected spring, rainwater; °includes connection to a piped sewer system, to a septictank, to a pit latrine, ventilated improved pit latrine, pit latrine with slab, composting toilet. Sharedimproved facilities are also included.

[page 334] [Geospatial Health 2019; 14:820]

Non-co

mmercial

use o

nly

DiscussionWe studied the determinants of the spatial pattern of height-

for-age at the household cluster level in Rwanda consideringdemographic, socio-economic variables and environmental vari-ables, with a proxy variable for aflatoxin exposure in children.Stunting prevalence in children in Rwanda was 30.2% which,according to the present stunting prevalence thresholds (de Onis et

al., 2019), is considered very high. Factors associated with stunt-ing were the child’s age, the mother’s height, the mother’s educa-tion, the child’s gender and its birth weight. The direction of therelationship for each covariate was as expected. First, the age ofchildren was negatively associated with height-for-age, a resultthat aligns with other studies in the sense that as a child becomesolder, the risk of stunting increases (Dewey and Huffman, 2009).Second, the mother’s height was positively associated with the

Article

Figure 4. Prevalence rate of Demographic and Health Survey variables per cluster. A) Prevalence rate of using deworming tablets in lastsix months; B) percentage of diarrhoea in last two weeks; C) non-improved water source in the households; D) non-improved sanitationin household.

Table 4. Regression coefficients of the socio-economic, environmental and accessibility factors on height-for-age.

Variables Non-standardized coefficients Standardized coefficients 95% CI B Standard error Beta Lower, Upper

Child’s age (months) -0.08** 0.01 -0.328 -0.102, -0.062Mother’s height (cm) 0.06** 0.01 0.223 0.036, 0.078Secondary education or higher 0.01* 0.00 0.128 0.002, 0.010Child being male -0.01* 0.00 -0.132 -0.010, -0.002Child’s birth weight (kg) 0.32* 0.13 0.102 0.065, 0.583Elevation (m) -0.00** 0.00 -0.207 -0.001, 0.000Cluster served by an urban vs a rural market -0.18* 0.09 -0.086 -0.351, -0.007B, Non-standardized coefficient; Beta, Standardized coefficient; CI, Confidence interval. *P≤0.05, **P≤0.001. Adjusted R2 of the model is 0.27.

[Geospatial Health 2019; 14:820] [page 335]

gh-2019_2.qxp_Hrev_master 12/11/19 13:53 Pagina 335

Non-co

mmercial

use o

nly

height-for-age of her child, showing the relation between maternalnutrition and anthropometric status of the child. Maternal shortstature is a high-risk factor of intra-uterine growth restriction,which has previously been linked to childhood stunting (Black etal., 2013). Third, we found that secondary education or higher ofthe mother was positively associated with height-for-age. InUganda, Wamani et al. (2007) found similar results. Chopra (2003)and Sakisaka et al. (2006) also identified that low maternal educa-tion was associated with lower height-for-age of children.

In our study, a child being male was negatively associated withheight-for-age. Previous Rwandan DHS surveys (INSR andMacro, 2006; NISR et al., 2010) have shown a similar pattern. TheRwanda DHS of 2010 showed that 47.4% of the boys were stuntedcompared to 41.4% of the girls. The Rwanda DHS of 2005 identi-fied that 46.3% of the boys were stunted compared to 44.4% of thegirls. In a study on sex differences in the nutritional status of HIV-exposed children in Rwanda, Condo et al. (2015) found significantdifferences in stunting status, with male children being more stunt-ed than their female counterparts, although they did not find a sig-nificant difference in feeding practices between male and femalechildren. As for birth weight, our findings showed that higher birthweight was positively associated with height-for-age. Black et al.(2013) have shown that children with normal birth weight are moreprotected against infections and have less mortality risk comparedto children with a low birth weight.

Although exclusive breastfeeding, use of deworming tablets,access to improved water source and sanitation of the householdare known to influence height-for-age (Stewart et al., 2013), ourmulti-variate analysis did not identify any significant association.Preceding birth interval, mothers’ BMI, and type of residence werealso not significantly associated with stunting. Households livingin areas of higher altitude negatively associated with the height-for-age of children. The robustness of the elevation covariate wasconsistent throughout the analysis. Similarly, Dang et al. (2004)found that children living in mountainous regions have a higherrisk of becoming stunted than children brought up in lowlands.Elevation could have a direct effect on height-for-age, but it ismore probable that these are more remote areas with higher foodinsecurity and lower levels of access to health care services. In acomprehensive vulnerability survey conducted in Rwanda in 2015,households at high altitudes were more at risk due to food insecu-rity, and their children were also found to be more stunted com-pared to those living in the rest of the country (MINAGRI et al.,2016). On the other hand, in this study we tested the accessibilityto health care services, and did not find a significant relationshipbetween distance to health facilities and height-for-age. This isprobably because the health care services delivery in Rwanda iswell organised with a universal health care system and thus equallyaccessible to the majority of the population (Binagwaho et al.,2014). We tested other environmental factors such as rainfall, rel-ative humidity and slope. These covariates were not statisticallysignificant. However, this non-significant result might result fromthe cross-sectional nature of our study. Considering seasonality infuture studies might reveal an impact of these factors on theanthropometric status of children. Skoufias and Vinha (2012) anal-ysed the impact of the weather on child height in rural Mexicofrom 1951 to 1985 and found that after positive rainfall shocks,defined as one or two standard deviations more than the average1951-1985 rainfall, children were shorter regardless of their loca-tion or altitude. Negative temperature shocks, defined as one ortwo standard deviations less than the average 1951-1985 tempera-

Article

Figure 5. Distribution of household clusters as served by the threecategories of markets. A) Within a 5-km distance; and B) a 10-kmdistance.

Figure 6. Moran’s I for the residuals.

[page 336] [Geospatial Health 2019; 14:820]

Non-co

mmercial

use o

nly

ture, also had a negative impact on child’s height but only for highaltitude areas (Skoufias and Vinha, 2012).

Household clusters that were served by rural markets had a sig-nificant negative association with height-for-age compared to clus-ters that were served by urban markets. This association had thesecond highest effect on the height-for-age of children after thechildren’s birth weight. This result supports our hypothesis thathouseholds that obtain food supplies from markets at the lower endof the food supply chain have a higher risk of aflatoxin exposure.Given that these findings are based on a proxy only, this findingshould be treated with caution. Also, in reality, the dynamics ofaflatoxin production along the food supply are complex whichmeans that our hypothesis requires further validation by studyingthe occurrence of this toxin along the food supply chain and quan-tifying contamination levels. On the other hand, due to the lack ofexplicit national spatial data on aflatoxins in foods, which is thecase in many developing countries, our approach could be appliedin other research settings on mycotoxins. Our study revealed thatthe identified variability in stunting prevalence at the cluster levelis lost when stunting levels are aggregated to the district level. Thiscan have considerable implications for stunting reduction policies

because a district might be classified as having low stunting andthus not be given priority for programme intervention even if highstunting prevalence could exist at the sub-district or cluster level.To enable more geographically targeted policies and programmesin future, it is important to consider cluster-level variations instunting prevalence.

The use of the Rwanda DHS 2014-15 data had severalstrengths and the consideration of biophysical factors and a proxymeasure of the exposure to aflatoxins in addition to socio-econom-ic factors rendered the analysis more robust, enabling spatiallyexplicit predictions of stunting occurrence. First, we analysed bothchild and mother characteristics, further increasing the robustnessof the analysis. Second, the DHS survey provided a range ofexplanatory variables of good quality data. Third, the availabilityof spatial data enabled us to analyse biophysical factors associatedwith stunting at the household cluster-level which, to the best ofour knowledge, has not been done before in Rwanda. Using thehousehold clusters, we could take into account the spatial variabil-ity of the covariates, visualise the spatial distribution of height-for-age at a fine-scale resolution, control for spatial autocorrelation,and finally make a spatial prediction map for height-for-age. Our

Article

[Geospatial Health 2019; 14:820] [page 337]

Figure 7. Model-predicted height-for-age z-scores (HAZ) per cluster. Based on Demographic and Health Survey, 2015. A) Per districtand B) with the same values shown in the background aggregated at the district level compared to the prevalence of stunting.

Non-co

mmercial

use o

nly

[page 338] [Geospatial Health 2019; 14:820]

study also had limitations. Although for the extraction of environ-mental factors and the estimation of household exposure to afla-toxins the displacement of the household clusters was taken intoaccount; there will be a margin of error as the true locations of theclusters remain unknown (Burgert et al., 2013). Also, our proxy foraflatoxins exposure via the food supply chain requires further evi-dence. This is because of the complex interplay of factors leadingto aflatoxins produced in food products (Sanchis and Magan, 2004;Sakisaka et al., 2006). Also, our scenario considered only the hier-archy of markets as the source of exposure, which does not takeinto account the exposure experienced by households that produceand store maize for their own consumption.

ConclusionsOur study confirms not only the usual effect of child and moth-

er factors on height-for-age but also shows the influence of envi-ronmental factors in determining the height-of-age of children inRwanda. Elevation and being supplied by markets at the lower endof the food supply chain were found to be significantly associatedwith low height-for-age. Thus, an understanding and considerationof the environmental drivers of stunting is crucial in order to pro-duce a holistic approach in addressing low height-for-age in chil-dren under five years. Our use of household clusters shows thevariability of stunting across the study area. In most publishedstudies, the analysis is generally conducted at the household levelwith no spatial component considered. Although this approach pro-vides valuable insights into the determinants of stunting on indi-vidual and household levels, most governmental interventions aretargeted at the regional scale. Future research should focus onstudying in depth the clustering observed in the height-for-agemeasure, to better understand the individual determinants of stunt-ing at a finer scale. Conducting such a research could shed light onoverlooked, yet determining correlates of stunting, which wouldresult in better geographically targeted interventions and prioritisa-tion of areas affected. It would also contribute to an examination ofthe temporal change in hotspots of stunting during the past years,which would lead to a better understanding of the spatial variationin the distribution of stunting. For Africa and Rwanda specifically,there is a need for more interdisciplinary research, incorporatinggeographical information system applications in understanding thecomplexity of stunting, to complement the research usually con-ducted on an individual or household level in the nutrition or socialscience fields. Finally, there is a tremendous data gap to be filledon the extent of mycotoxin exposure on the national level and howthis affects linear growth, not only at the individual level but alsoat a regional and national scale.

ReferencesAmuzie C, Bandyopadhyay R, Bhat RV, Black R, Burger H,

Cardwell KF, Gelderblom W, Gong YY, Groopman JD,Kimanya M, Miller JD, Oswald I, Pascale M, Payne GA,Phillips TD, Riley R, Shephard GS, Stoltzfus R, Sugita-Konishi Y, Turner PC, Wogan GN, Wu F, 2016. Mycotoxincontrol in low- and middle-income countries. Wild CP, MillerD, Groopman JD, Lyon, France.

AUC, NEPAD, 2013. The cost of hunger in Rwanda, social and

economic impacts of child undernutrition in Rwanda, implica-tions on national development and vision 2020. African UnionCommission (ACU), NEPAD Planning and CoordinatingAgency, Economic Commission for Africa (ECA), World FoodProgramme (WFP), Rwanda.

Binagwaho A, Farmer PE, Nsanzimana S, Karema C, Gasana M,de Dieu Ngirabega J, Ngabo F, Wagner CM, Nutt CT, NyatanyiT, Gatera M, Kayiteshonga Y, Mugeni C, Mugwaneza P,Shema J, Uwaliraye P, Gaju E, Muhimpundu MA, Dushime T,Senyana F, Mazarati JB, Gaju CM, Tuyisenge L, Mutabazi V,Kyamanywa P, Rusanganwa V, Nyemazi JP, Umutoni A,Kankindi I, Ntizimira C, Ruton H, Mugume N, Nkunda D,Ndenga E, Mubiligi JM, Kakoma JB, Karita E, Sekabaraga C,Rusingiza E, Rich ML, Mukherjee JS, Rhatigan J, CanceddaC, Bertrand-Farmer D, Bukhman G, Stulac SN, Tapela NM,van der Hoof Holstein C, Shulman LN, Habinshuti A, BondsMH, Wilkes MS, Lu C, Smith-Fawzi MC, Swain JD, MurphyMP, Ricks A, Kerry VB, Bush BP, Siegler RW, Stern CS,Sliney A, Nuthulaganti T, Karangwa I, Pegurri E, Dahl O,Drobac PC, 2014. Rwanda 20 years on: investing in life.Lancet 384:371-5.

Black RE, Allen LH, Bhutta ZA, Caulfield LE, de Onis M, EzzatiM, Mathers C, Rivera J, Maternal, Child Undernutrition StudyG, 2008. Maternal and child undernutrition: global and region-al exposures and health consequences. Lancet 371:243-60.

Black RE, Victora CG, Walker SP, Bhutta ZA, Christian P, de OnisM, Ezzati M, Grantham-McGregor S, Katz J, Martorell R,Uauy R, Maternal, Child Nutrition Study G, 2013. Maternaland child undernutrition and overweight in low-income andmiddle-income countries. Lancet 382:427-51.

Burgert CR, Colston J, Roy T, Zachary B, 2013. Geographic dis-placement procedure and georeferenced data release policy forthe Demographic and Health Surveys. DHS Spatial AnalysisReports No. 7. ICF International, Calverton, Maryland, USA.

Chopra M, 2003. Risk factors for undernutrition of young childrenin a rural area of South Africa. Public Health Nutr 6:645-52.

Condo JU, Gage A, Mock N, Rice J, Greiner T, 2015. Sex differ-ences in nutritional status of HIV-exposed children in Rwanda:a longitudinal study. Trop Med Int Health 20:17-23.

Cotty PJ, Jaime-Garcia R, 2007. Influences of climate on aflatoxinproducing fungi and aflatoxin contamination. Int J FoodMicrobiol 119:109-15.

Dang S, Yan H, Yamamoto S, Wang X, Zeng L, 2004. Poor nutri-tional status of younger Tibetan children living at high alti-tudes. Eur J Clin Nutr 58:938-46.

de Onis M, Borghi E, Arimond M, Webb P, Croft T, Saha K, De-Regil LM, Thuita F, Heidkamp R, Krasevec J, Hayashi C,Flores-Ayala R, 2019. Prevalence thresholds for wasting, over-weight and stunting in children under 5 years. Public HealthNutr 22:175-9.

de Onis M, Branca F, 2016. Childhood stunting: a global perspec-tive. Matern Child Nutr 12:12-26.

de Sherbinin A, 2011. The biophysical and geographical correlatesof child malnutrition in Africa. Popul Space Place 17:27-46.

Dewey KG, Begum K, 2011. Long-term consequences of stuntingin early life. Matern Child Nutr 7:5-18.

Dewey KG, Huffman SL, 2009. Maternal, infant, and young childnutrition: combining efforts to maximize impacts on childgrowth and micronutrient status. Food Nutr Bull 30:S187-9.

FAO, IFAD, UNICEF, WHFP, WHO, 2017. The State of FoodSecurity and Nutrition in the World 2017. Building resilience

Article

Non-co

mmercial

use o

nly

for peace and food security. FAO, Rome, Italy.Gbashi S, Madala NE, Adekoya I, Adebo O, De Saeger S, De

Boevre M, Njobeh PB, 2018. The socio-economic impact ofmycotoxin contamination in Africa. In Njobeh PB, ed. Fungiand mycotoxins - their occurrence, impact on health and theeconomy as well as pre- and postharvest management strate-gies. Available from: https://www.intechopen.com/books/mycotoxins-impact-and-management-strategies/the-socio-economic-impact-of-mycotoxin-contamination-in-africa

Hell K, Mutegi C, 2011. Aflatoxin control and prevention strate-gies in key crops of Sub-Saharan Africa. Afr J Microbiol Res5:459-66.

INSR, Macro O, 2006. Rwanda Demographic and Health Survey2005. Institut National de la Statistique du Rwanda (INSR),ORC Macro, Calverton, Maryland, USA.

Kamika I, Ngbolua KTN, Tekere M, 2016. Occurrence of aflatoxincontamination in maize throughout the supply chain in theDemocratic Republic of Congo. Food Control 69:292-6.

Khlangwiset P, Shephard GS, Wu F, 2011. Aflatoxins and growthimpairment: a review. Crit Rev Toxicol 41:740-55.

Lu C, Mejia-Guevara I, Hill K, Farmer P, Subramanian SV,Binagwaho A, 2016. Community-based health financing andchild stunting in rural Rwanda. Am J Public Health 106:49-55.

Mendelsohn J, Jarvis A, Robertson T, Nyiratuza M, 2016. Rwanda:the measure of a land. Vital Signs, Kigali, Rwanda.

MIGEPROF, 2018. Rwanda Country Strategic Review of Foodand Nutrition Security. Ministry of Gender and FamilyPromotion, Kigali, Rwanda.

MINAGRI, NISR, WFP, 2016. Rwanda 2015-Comprehensive foodsecurity and vulnerability analysis. MINAGRI, NISR, WFP,Kigali, Rwanda.

MINALOC, MOH, MINAGRI, 2014. National Food and NutritionStrategic Plan 2013-2018. MINALOC, MOH, MINAGRI,Rwanda.

Nikoi E, Anthamatten P, 2013. An examination of environmentalcorrelates with childhood height-for-age in Ghana. PublicHealth Nutr 16:46-53.

NISR, MINECOFIN, 2012. Fourth Rwanda Population andHousing Census (2012 RPHC) Census Atlas. National Instituteof Statistics, Ministry of Finance and Economic Planning,Rwanda.

NISR, MOH, ICF, 2010. Rwanda Demographic and Health Survey2010. National Institute of Statistics of Rwanda (NISR),Ministry of Health (MOH), ICF International, Calverton,Maryland, USA.

NISR, MOH, ICF, 2015. Rwanda Demographic and Health Survey2014-15. National Institute of Statistics of Rwanda (NISR),Ministry of Health (MOH), ICF International, Rockville,Maryland, USA.

Nyandwi E, Veldkamp T, Amer S, 2016. Regional climate sensitiv-ity of wetland environments in Rwanda: the need for a loca-tion-specific approach. Reg Environ Change 16:1635-47.

Onis M, 2007. WHO Child Growth Standards based onlength/height, weight and age. Acta Paediatr 95:76-85.

Pitt JI, 2000. Toxigenic fungi and mycotoxins. Br Med Bull56:184-92.

Rutstein SO, Rojas G, 2006. Guide to DHS Statistics Demographicand Health Surveys Methodology. Measure DHS/ICFInternational, Calverton, Maryland, USA.

Sakisaka K, Wakai S, Kuroiwa C, Flores LC, Kai I, MercedesArago’n M, Hanada K, 2006. Nutritional status and associatedfactors in children aged 0–23 months in Granada, Nicaragua.Public Health 120:400-11.

Sanchis V, Magan N, 2004. 8 - Environmental conditions affectingmycotoxins. In: Magan N, Olsen M, eds. Mycotoxins in food.Woodhead Publishing, Cambridge, UK. pp 174-189.

Shephard GS, 2008. Impact of mycotoxins on human health indeveloping countries. Food Addit Contam Part A Chem AnalControl Expo Risk Assess 25:146-51.

Skoufias E, Vinha K, 2012. Climate variability and child height inrural Mexico. Econ Hum Biol 10:54-73.

Stewart CP, Iannotti L, Dewey KG, Michaelsen KF, Onyango AW,2013. Contextualising complementary feeding in a broaderframework for stunting prevention. Matern Child Nutr 9:27-45.

Udomkun P, Wiredu AN, Nagle M, Bandyopadhyay R, Muller J,Vanlauwe B, 2017. Mycotoxins in Sub-Saharan Africa: Presentsituation, socio-economic impact, awareness, and outlook.Food Control 72:110-22.

Uwiringiyimana V, Ocke MC, Amer S, Veldkamp A, 2018.Predictors of stunting with particular focus on complementaryfeeding practices: A cross-sectional study in the northernprovince of Rwanda. Nutrition 60:11-8.

Victora CG, Adair L, Fall C, Hallal PC, Martorell R, Richter L,Sachdev HS, Maternal, Child Undernutrition Study G, 2008.Maternal and child undernutrition: consequences for adulthealth and human capital. Lancet 371:340-57.

Victora CG, de Onis M, Hallal PC, Blossner M, Shrimpton R,2010. Worldwide timing of growth faltering: revisiting impli-cations for interventions. Pediatrics 125:e473-80.

Vilcins D, Sly PD, Jagals P, 2018. Environmental risk factors asso-ciated with child stunting: a systematic review of the literature.Ann Global Health 84:551.

Voss PR, Long DD, Hammer RB, Friedman S, 2006. County childpoverty rates in the US: a spatial regression approach. PopulRes Policy Rev 25:369-91.

Wamani H, Åstrøm AN, Peterson S, Tumwine JK, Tylleskär T,2007. Predictors of poor anthropometric status among childrenunder 2 years of age in rural Uganda. Public Health Nutr9:320-6.

Wild CP, Gong YY, 2010. Mycotoxins and human disease: a large-ly ignored global health issue. Carcinogenesis 31:71-82.

Article

[Geospatial Health 2019; 14:820] [page 339]

Non-co

mmercial

use o

nly